Currently the handle type is a global pointer which holds 8 bytes.
We need a larger type which hold 16 bytes, therefore change it
to [i64 x 2].
Differential Revision: https://reviews.llvm.org/D48094
llvm-svn: 334625
- Do not emit following assembler directives:
- .hsa_code_object_version
- .hsa_code_object_isa
- .amd_amdgpu_isa
- .amd_amdgpu_hsa_metadata
- .amd_amdgpu_pal_metadata
- Do not emit .note entries
- Cleanup and bring in sync kernel descriptor header file
- Emit kernel descriptor into .rodata with appropriate relocations and
alignments
llvm-svn: 334519
The use iterator, used within findMaskOperands(), can return anything which is
not a def. isUse() requires a register, so check isReg() before calling isUse().
Differential Revision: https://reviews.llvm.org/D48047
llvm-svn: 334459
Rational: if there is indirect access that is usually an issue
because load is not ready by the use. However, if use is inside a
loop and load is outside that is potentially an issue for a first
iteration only.
Differential Revision: https://reviews.llvm.org/D47740
llvm-svn: 334420
Summary:
The idiom recognition seems rather poor.
Only the `@bzhi32_d0` produces `v_bfe_u32`.
But they all should.
This needs to be fixed before D47980 can be re-landed.
Reviewers: mareko, bogner, rampitec, arsenm, tstellar, nhaehnle
Reviewed By: nhaehnle
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Tags: #amdgpu
Differential Revision: https://reviews.llvm.org/D48005
llvm-svn: 334398
AMDGPU inline assembler support i16, half and i128 typed variables in constraints, but they were reported as error.
Needed to fix https://github.com/RadeonOpenCompute/ROCm/issues/341,
e.g. to be able to load with global_load_dwordx4 to a 128bit integer variable
Differential Revision: https://reviews.llvm.org/D44920
llvm-svn: 334301
- Make code easier to maintain.
- Avoid generating waitcnts for VMEM if the address sppace does not involve VMEM.
- Add support to generate waitcnts for LDS and GDS memory.
Differential Revision: https://reviews.llvm.org/D47504
llvm-svn: 334241
This has two main components. First, widen
widen short constant loads in DAG when they have
the correct alignment. This is already done a bit in
AMDGPUCodeGenPrepare, since that has access to
DivergenceAnalysis. This can't help kernarg loads
created in the DAG. Start to use DAG divergence analysis
to help this case.
The second part is to avoid kernel argument lowering
breaking the alignment of short vector elements because
calling convention lowering wants to split everything
into legal register types.
When loading a split type, load the nearest 4-byte aligned
segment and shift to get the desired bits. This extra
load of the earlier argument piece ends up merging,
and the bit extract hopefully folds out.
There are a number of improvements and regressions with
this, but I think as-is this is a better compromise between
several of the worst parts of SelectionDAG.
Particularly when i16 is legal, this produces worse code
for i8 and i16 element vector kernel arguments. This is
partially due to the very weak load merging the DAG does.
It only looks for fairly specific combines between pairs
of loads which no longer appear. In particular this
causes v4i16 loads to be split into 2 components when
previously the two halves were merged.
Worse, because of the newly introduced shifts, there
is a lot more unnecessary vector packing and unpacking code
emitted. At least some of this is due to reporting
false for isTypeDesirableForOp for i16 as a workaround for
the lack of divergence information in the DAG. The cases
where this happens it doesn't actually matter, but the
relevant code in SimplifyDemandedBits doens't have the context
to know to ignore this.
The use of the scalar cache is probably more important
than the mess of mostly scalar instructions doing this packing
and unpacking. Future work can fix this, possibly by making better
use of the new DAG divergence information for controlling promotion
decisions, or adding another version of shift + trunc + shift
combines that doesn't only know about the used types.
llvm-svn: 334180
When denormals are supported we are producing a full division for
1.0f / x. That still can be replaced by the faster version:
bool c = fabs(x) > 0x1.0p+96f;
float s = c ? 0x1.0p-32f : 1.0f;
x *= s;
return s * v_rcp_f32(x)
in case if requested accuracy is 2.5ulp or less. The same version
is used if denormals are not supported for non 1.0 numerators, where
just v_rcp_f32 is then used for 1.0 numerator.
The optimization of 1/x is extended to the case -1/x, which is the
same except for the resulting sign bit.
OpenCL conformance passed with both enabled and disabled denorms.
Differential Revision: https://reviews.llvm.org/D47805
llvm-svn: 334142
Fixes terrible code on targets without f16 support. The
legalization creates a mess that is difficult to recover
from. Also should avoid randomly breaking these tests
multiple times in sequence in future commits.
Some regressions in cases where it happens to be better
to pull the source modifier after the conversion.
llvm-svn: 334132
Preserves the low bound of the !range. I don't think
it's legal to do anything with the top half since it's
theoretically reading garbage.
llvm-svn: 334045
When legalizing illegal FP load results, this was
for some reason dropping the invariant and dereferencable
memory flags. There doesn't seem to be any reason for this,
and the equivalent isn't done for integer loads.
Fixes an issue in a future AMDGPU commit where some identical
loads fail to merge because one of the loads ends up
dropping the flags.
llvm-svn: 334020
On GFX9 and earlier, flat memory ops may decrement VMCNT out-of-order as well as LGKMCNT out-of-order.
Differential Revision: https://reviews.llvm.org/D46616
llvm-svn: 333926
Memory clauses are formed into bundles in presence of xnack.
Their source operands are marked as early-clobber.
This allows to allocate distinct source and destination registers
within a clause and prevent breaking the clause with s_nop in the
hazard recognizer.
Clauses are undone before post-RA scheduler to allow some rescheduling,
which will not break the clause since artificial edges are created in
the dag to keep memory operations together. Yet this allows a better
ILP in some cases.
Differential Revision: https://reviews.llvm.org/D47511
llvm-svn: 333691
v2: use "ensureAlignment"
make functions cache line aligned
Fixes GPU hangs since r333219:
"AMDGPU: Split R600 AsmPrinter code into its own class"
Differential Revision: https://reviews.llvm.org/D47516
llvm-svn: 333622
This was just emitting loads with the ABI alignment
for the raw type. The true alignment is often better,
especially when an illegal vector type was scalarized.
The better alignment allows using a scalar load
more often.
llvm-svn: 333558
In terms of waitcnt insertion/if necessary, the waitcnt pass forces convergence
for a loop. Previously, that kicked if greater than 2 passes over a loop, which
doesn't account for loop with many bottom blocks. So, increase the threshold to
(n+1), where n is the number of bottom blocks. This gives the pass an
opportunity to consider the contribution of each bottom block, to the overall
loop, before the forced convergence potentially kicks in.
Differential Revision: https://reviews.llvm.org/D47488
llvm-svn: 333556
AFAIK the driver's allocation will actually have to round this
up anyway. It is useful to track the rounded up size, so that
the end of the kernel segment is known to be dereferencable so
a wider s_load_dword can be used for a short argument at the end
of the segment.
llvm-svn: 333456
Summary:
For a block with WQM on entry and exit and containing no exact mode
code, but containing some WWM code, the WQM pass forgot to process the
block at all and so did not insert code to enter and leave WWM.
This commit fixes that.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D47027
Change-Id: I044792eead1293bed4203fb26ce75f47878afeb6
llvm-svn: 333362
With the removal of the old waitcnt pass, the '-enable-si-insert-waitcnts' option is obsolete. Remove it.
Differential Revision: https://reviews.llvm.org/D47378
llvm-svn: 333303
This is adoption of HSAIL perfhint pass. Two types of hints are produced:
1. Function is memory bound.
2. Kernel can use wave limiter.
Currently these hints are used in the scheduler. If a function is suspected
to be memory bound we allow occupancy to decrease to 4 waves in the course
of scheduling.
Differential Revision: https://reviews.llvm.org/D46992
llvm-svn: 333289
Summary:
Lower control flow did not correctly handle the case that a loop break
in if/else was on a condition that was not guaranteed to be masked by
exec. The first test kernel shows an example of this going wrong; after
exiting the loop, exec is all ones, even if it was not before the loop.
The fix is for lowering of if-break and else-break to insert an
S_AND_B64 to mask the break condition with exec. This commit also
includes the optimization of not inserting that S_AND_B64 if it is
obviously not needed because the break condition is the result of a
V_CMP in the same basic block.
V2: Addressed some review comments.
V3: Test fixes.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D44046
Change-Id: I0fc56a01209a9e99d1d5c9b0ffd16f111caf200c
llvm-svn: 333258
Summary:
StructurizeCFG::orderNodes basically uses a reverse post-order (RPO) traversal of the region list to get the order.
The only problem with it is that sometimes backedges for outer loops will be visited before backedges for inner loops.
To solve this problem, a loop depth based approach has been used to make sure all blocks in this loop has been visited
before moving on to outer loop.
However, we found a problem for a SubRegion which is a loop itself:
--> BB1 --> BB2 --> BB3 -->
In this case, BB2 is a SubRegion (loop), and thus its loopdepth is different than that of BB1 and BB3. This fact will lead
BB2 to be placed in the wrong order.
In this work, we treat the SubRegion as a special case and use its exit block to determine the loop and its depth
to guard the sorting.
Reviewers:
arsenm, jlebar
Differential Revision:
https://reviews.llvm.org/D46912
llvm-svn: 333111
The integer operation convertion for some reason only happens
if the source is a bitcast from an integer, which happens to
always be the situation when the result is loaded. Add
an additional pattern for when the source operation is really
an FP operation.
llvm-svn: 333019
This usually results in better code. Fixes using
inline asm with short2, and also fixes having a different
ABI for function parameters between VI and gfx9.
Partially cleans up the mess used for lowering of the d16
operations. Making v4f16 legal will help clean this up more,
but this requires additional work.
llvm-svn: 332953
This is the FP sibling of D43141 with the corresponding IR change in rL327212.
We can't propagate undef here because if a variable operand is a NaN, these
binops must propagate NaN. Neither global nor node-level fast-math makes a
difference. If we have 'nnan', I think later folds can turn the NaN into undef.
The tests in X86/fp-undef.ll are meant to be the definitive verification for
these folds - everything reduces identically now.
The other test changes are collateral damage. They may need to be altered to
preserve their intent.
Differential Revision: https://reviews.llvm.org/D47026
llvm-svn: 332920
Eliminate loads from the dispatch packet when they will have
a known value.
Also pattern match the code used by the library to handle partial
workgroup dispatches, which isn't necessary if reqd_work_group_size
is used.
llvm-svn: 332771
Summary:
The current StructurizeCFG pass only works for CFG with one exit. AMDGPUUnifyDivergentExitNodes combines multiple "return" blocks and/or "unreachable" blocks
to one exit block for the Structurizer to work. However, infinite loop is another kind of special "exit", and if we don't handle it, the case of multiple exits will prevent the structurizer from working.
In this work, for each infinite loop, we add a dummy edge to the "return" block, and thus the AMDGPUUnifyDivergentExitNodes pass will work with infinite loops.
This will make CFG with infinite loops be structurized.
Reviewer:
nhaehnle
Differential Revision:
https://reviews.llvm.org/D46340
llvm-svn: 332625
It is legal for the type passed to isLegalAddressingMode to be
unsized or, more specifically, VoidTy. In this case, we must
check the legality of load / stores for all legal types. Directly
trying to call getTypeStoreSize is incorrect, and leads to breakage
in e.g. Loop Strength Reduction. This change guards against that
behaviour.
Differential Revision: https://reviews.llvm.org/D40405
llvm-svn: 332409
